Holistic Evaluation of Language Models

Language models (LMs) are becoming the foundation for almost all major language technologies, but their capabilities, limitations, and risks are not well understood. We present Holistic Evaluation of Language Models (HELM) to improve the transparency of language models. First, we taxonomize the vast space of potential scenarios (i.e. use cases) and metrics (i.e. desiderata) that are of interest for LMs. Then we select a broad subset based on coverage and feasibility, noting what's missing or underrepresented (e.g. question answering for neglected English dialects, metrics for trustworthiness). Second, we adopt a multi-metric approach: We measure 7 metrics (accuracy, calibration, robustness, fairness, bias, toxicity, and efficiency) for each of 16 core scenarios when possible (87.5% of the time). This ensures metrics beyond accuracy don't fall to the wayside, and that trade-offs are clearly exposed. We also perform 7 targeted evaluations, based on 26 targeted scenarios, to analyze specific aspects (e.g. reasoning, disinformation). Third, we conduct a large-scale evaluation of 30 prominent language models (spanning open, limited-access, and closed models) on all 42 scenarios, 21 of which were not previously used in mainstream LM evaluation. Prior to HELM, models on average were evaluated on just 17.9% of the core HELM scenarios, with some prominent models not sharing a single scenario in common. We improve this to 96.0%: now all 30 models have been densely benchmarked on the same core scenarios and metrics under standardized conditions. Our evaluation surfaces 25 top-level findings. For full transparency, we release all raw model prompts and completions publicly for further analysis, as well as a general modular toolkit. We intend for HELM to be a living benchmark for the community, continuously updated with new scenarios, metrics, and models.

LAD: Language Augmented Diffusion for Reinforcement Learning

Learning skills from language provides a powerful avenue for generalization in reinforcement learning, although it remains a challenging task as it requires agents to capture the complex interdependencies between language, actions, and states. In this paper, we propose leveraging Language Augmented Diffusion models as a planner conditioned on language (LAD). We demonstrate the comparable performance of LAD with the state-of-the-art on the CALVIN language robotics benchmark with a much simpler architecture that contains no inductive biases specialized to robotics, achieving an average success rate (SR) of 72% compared to the best performance of 76%. We also conduct an analysis on the properties of language conditioned diffusion in reinforcement learning.

DecAF: Joint Decoding of Answers and Logical Forms for Question Answering over Knowledge Bases

Question answering over knowledge bases (KBs) aims to answer natural language questions with factual information such as entities and relations in KBs. Previous methods either generate logical forms that can be executed over KBs to obtain final answers or predict answers directly. Empirical results show that the former often produces more accurate answers, but it suffers from non-execution issues due to potential syntactic and semantic errors in the generated logical forms. In this work, we propose a novel framework DecAF that jointly generates both logical forms and direct answers, and then combines the merits of them to get the final answers. Moreover, different from most of the previous methods, DecAF is based on simple free-text retrieval without relying on any entity linking tools -- this simplification eases its adaptation to different datasets. DecAF achieves new state-of-the-art accuracy on WebQSP, FreebaseQA, and GrailQA benchmarks, while getting competitive results on the ComplexWebQuestions benchmark.

Emotion Recognition in Conversation using Probabilistic Soft Logic

Creating agents that can both appropriately respond to conversations and understand complex human linguistic tendencies and social cues has been a long standing challenge in the NLP community. A recent pillar of research revolves around emotion recognition in conversation (ERC); a sub-field of emotion recognition that focuses on conversations or dialogues that contain two or more utterances. In this work, we explore an approach to ERC that exploits the use of neural embeddings along with complex structures in dialogues. We implement our approach in a framework called Probabilistic Soft Logic (PSL), a declarative templating language that uses first-order like logical rules, that when combined with data, define a particular class of graphical model. Additionally, PSL provides functionality for the incorporation of results from neural models into PSL models. This allows our model to take advantage of advanced neural methods, such as sentence embeddings, and logical reasoning over the structure of a dialogue. We compare our method with state-of-the-art purely neural ERC systems, and see almost a 20% improvement. With these results, we provide an extensive qualitative and quantitative analysis over the DailyDialog conversation dataset.

NeuPSL: Neural Probabilistic Soft Logic

We present Neural Probabilistic Soft Logic (NeuPSL), a novel neuro-symbolic (NeSy) framework that unites state-of-the-art symbolic reasoning with the low-level perception of deep neural networks. To explicitly model the boundary between neural and symbolic representations, we introduce NeSy Energy-Based Models, a general family of energy-based models that combine neural and symbolic reasoning. Using this framework, we show how to seamlessly integrate neural and symbolic parameter learning and inference. We perform an extensive empirical evaluation and show that NeuPSL outperforms existing methods on joint inference and has significantly lower variance in almost all settings.

On the Opportunities and Risks of Foundation Models

AI is undergoing a paradigm shift with the rise of models (e.g., BERT, DALL-E, GPT-3) that are trained on broad data at scale and are adaptable to a wide range of downstream tasks. We call these models foundation models to underscore their critically central yet incomplete character. This report provides a thorough account of the opportunities and risks of foundation models, ranging from their capabilities (e.g., language, vision, robotics, reasoning, human interaction) and technical principles(e.g., model architectures, training procedures, data, systems, security, evaluation, theory) to their applications (e.g., law, healthcare, education) and societal impact (e.g., inequity, misuse, economic and environmental impact, legal and ethical considerations). Though foundation models are based on standard deep learning and transfer learning, their scale results in new emergent capabilities,and their effectiveness across so many tasks incentivizes homogenization. Homogenization provides powerful leverage but demands caution, as the defects of the foundation model are inherited by all the adapted models downstream. Despite the impending widespread deployment of foundation models, we currently lack a clear understanding of how they work, when they fail, and what they are even capable of due to their emergent properties. To tackle these questions, we believe much of the critical research on foundation models will require deep interdisciplinary collaboration commensurate with their fundamentally sociotechnical nature.

FoveaTer: Foveated Transformer for Image Classification

Many animals and humans process the visual field with a varying spatial resolution (foveated vision) and use peripheral processing to make eye movements and point the fovea to acquire high-resolution information about objects of interest. This architecture results in computationally efficient rapid scene exploration. Recent progress in vision Transformers has brought about new alternatives to the traditionally convolution-reliant computer vision systems. However, these models do not explicitly model the foveated properties of the visual system nor the interaction between eye movements and the classification task. We propose foveated Transformer (FoveaTer) model, which uses pooling regions and saccadic movements to perform object classification tasks using a vision Transformer architecture. Our proposed model pools the image features using squared pooling regions, an approximation to the biologically-inspired foveated architecture, and uses the pooled features as an input to a Transformer Network. It decides on the following fixation location based on the attention assigned by the Transformer to various locations from previous and present fixations. The model uses a confidence threshold to stop scene exploration, allowing to dynamically allocate more fixation/computational resources to more challenging images. We construct an ensemble model using our proposed model and unfoveated model, achieving an accuracy 1.36% below the unfoveated model with 22% computational savings. Finally, we demonstrate our model's robustness against adversarial attacks, where it outperforms the unfoveated model.

Open Question Answering over Tables and Text

In open question answering (QA), the answer to a question is produced by retrieving and then analyzing documents that might contain answers to the question. Most open QA systems have considered only retrieving information from unstructured text. Here we consider for the first time open QA over both tabular and textual data and present a new large-scale dataset Open Table-Text Question Answering (OTT-QA) to evaluate performance on this task. Most questions in OTT-QA require multi-hop inference across tabular data and unstructured text, and the evidence required to answer a question can be distributed in different ways over these two types of input, making evidence retrieval challenging---our baseline model using an iterative retriever and BERT-based reader achieves an exact match score less than 10%. We then propose two novel techniques to address the challenge of retrieving and aggregating evidence for OTT-QA. The first technique is to use "early fusion" to group multiple highly relevant tabular and textual units into a fused block, which provides more context for the retriever to search for. The second technique is to use a cross-block reader to model the cross-dependency between multiple retrieved evidences with global-local sparse attention. Combining these two techniques improves the score significantly, to above 27%.

GAT-GMM: Generative Adversarial Training for Gaussian Mixture Models

Generative adversarial networks (GANs) learn the distribution of observed samples through a zero-sum game between two machine players, a generator and a discriminator. While GANs achieve great success in learning the complex distribution of image, sound, and text data, they perform suboptimally in learning multi-modal distribution-learning benchmarks including Gaussian mixture models (GMMs). In this paper, we propose Generative Adversarial Training for Gaussian Mixture Models (GAT-GMM), a minimax GAN framework for learning GMMs. Motivated by optimal transport theory, we design the zero-sum game in GAT-GMM using a random linear generator and a softmax-based quadratic discriminator architecture, which leads to a non-convex concave minimax optimization problem. We show that a Gradient Descent Ascent (GDA) method converges to an approximate stationary minimax point of the GAT-GMM optimization problem. In the benchmark case of a mixture of two symmetric, well-separated Gaussians, we further show this stationary point recovers the true parameters of the underlying GMM. We numerically support our theoretical findings by performing several experiments, which demonstrate that GAT-GMM can perform as well as the expectation-maximization algorithm in learning mixtures of two Gaussians.

HybridQA: A Dataset of Multi-Hop Question Answering over Tabular and Textual Data

Existing question answering datasets focus on dealing with homogeneous information, based either only on text or KB/Table information alone. However, as human knowledge is distributed over heterogeneous forms, using homogeneous information might lead to severe coverage problems. To fill in the gap, we present \dataset, a new large-scale question-answering dataset that requires reasoning on heterogeneous information. Each question is aligned with a structured Wikipedia table and multiple free-form corpora linked with the entities in the table. The questions are designed to aggregate both tabular information and text information, i.e. lack of either form would render the question unanswerable. We test with three different models: 1) table-only model. 2) text-only model. 3) a hybrid model \model which combines both table and textual information to build a reasoning path towards the answer. The experimental results show that the first two baselines obtain compromised scores below 20\%, while \model significantly boosts EM score to over 50\%, which proves the necessity to aggregate both structure and unstructured information in \dataset. However, \model's score is still far behind human performance, hence we believe \dataset to an ideal and challenging benchmark to study question answering under heterogeneous information. The dataset and code are available at \url{https://github.com/wenhuchen/HybridQA}.